To determine the risk levels of urban signal-controlled intersection lanes, we propose two comprehensive risk assessment sets that incorporate time proximity, spatial proximity, and velocity-based indicators. Using actual trajectory data of intersection users, we analyse the interactions between three common vulnerable road users and motor vehicles. Bayesian inference algorithm is employed to evaluate lane risk levels. Additionally, XGBoost, random algorithm, and SVM algorithm with a Gaussian kernel are utilized to predict intersection lane risk. The importance and correlation of risk factors are explored using SHAP value theory. Results indicate that while extremely dangerous collisions are less likely due to signal control, moderate risk remains prevalent. The XGBoost model outperforms the other methods with an accuracy rate of 91.51% and precision rate of 94.17%. Among the analysed interaction risk factors, the speed and acceleration of road users have the greatest impact on lane risk levels. In low traffic flow conditions, higher speeds of motorcycles and cars increase collision risks. Lanes three and four exhibit higher probabilities of hazardous interactions between pedestrians and cars, resulting in more severe pedestrian injuries. Additionally, motorcycles at lower speeds are more prone to risky interactions with pedestrians.